import argparse import logging import torch import torch.nn as nn from torch import cuda from torch.autograd import Variable from torch.utils.data import DataLoader,Dataset import torchvision import torchvision.datasets as dset import torchvision.transforms as transforms import torchvision.utils from PIL import Image import torch.nn.functional as F import matplotlib.pyplot as plt import numpy as np import random from custom_transform import CustomResize from custom_transform import CustomToTensor from AD_Standard_CNN_Dataset import AD_Standard_CNN_Dataset from cnn_3d_with_ae import CNN logging.basicConfig( format='%(asctime)s %(levelname)s: %(message)s', datefmt='%Y-%m-%d %H:%M:%S', level=logging.INFO) parser = argparse.ArgumentParser(description="Starter code for CNN .") parser.add_argument("--epochs", default=20, type=int, help="Epochs through the data. (default=20)") parser.add_argument("--learning_rate", "-lr", default=1e-3, type=float, help="Learning rate of the optimization. (default=0.01)") parser.add_argument('--weight_decay', '--wd', default=1e-4, type=float, metavar='W', help='weight decay (default: 1e-4)') parser.add_argument("--batch_size", default=1, type=int, help="Batch size for training. (default=1)") parser.add_argument("--gpuid", default=[0], nargs='+', type=int, help="ID of gpu device to use. Empty implies cpu usage.") parser.add_argument("--autoencoder", default=True, type=bool, help="Whether to use the parameters from pretrained autoencoder.") parser.add_argument("--num_classes", default=2, type=int, help="The number of classes, 2 or 3.") parser.add_argument("--estop", default=1e-5, type=float, help="Early stopping criteria on the development set. (default=1e-2)") parser.add_argument("--noise", default=True, type=bool, help="Whether to add gaussian noise to scans.") # feel free to add more arguments as you need def main(options): # Path configuration if options.num_classes == 2: TRAINING_PATH = 'train_2C_new.txt' TESTING_PATH = 'validation_2C_new.txt' else: TRAINING_PATH = 'train.txt' TESTING_PATH = 'test.txt' IMG_PATH = './NewWhole' trg_size = (121, 145, 121) # transformations = transforms.Compose([CustomResize("CNN", trg_size), # CustomToTensor("CNN") # ]) dset_train = AD_Standard_CNN_Dataset(IMG_PATH, TRAINING_PATH, noise=True) dset_test = AD_Standard_CNN_Dataset(IMG_PATH, TESTING_PATH, noise=False) # Use argument load to distinguish training and testing train_loader = DataLoader(dset_train, batch_size = options.batch_size, shuffle = True, num_workers = 4, drop_last = True ) test_loader = DataLoader(dset_test, batch_size = options.batch_size, shuffle = False, num_workers = 4, drop_last=True ) use_cuda = (len(options.gpuid) >= 1) # if options.gpuid: # cuda.set_device(options.gpuid[0]) # Training process model = CNN(options.num_classes) if use_cuda > 0: model = model.cuda() else: model.cpu() if options.autoencoder: pretrained_ae = torch.load("./autoencoder_pretrained_model39") model.state_dict()['conv1.weight'] = pretrained_ae['encoder.weight'].view(410,1,7,7,7) model.state_dict()['conv1.bias'] = pretrained_ae['encoder.bias'] for p in model.conv1.parameters(): p.requires_grad = False criterion = torch.nn.NLLLoss() lr = options.learning_rate optimizer = torch.optim.Adam(filter(lambda x: x.requires_grad, model.parameters()), lr, weight_decay=options.weight_decay) # main training loop last_dev_loss = 1e-4 max_acc = 0 max_epoch = 0 f1 = open("cnn_autoencoder_loss_train", 'a') f2 = open("cnn_autoencoder_loss_dev", 'a') for epoch_i in range(options.epochs): logging.info("At {0}-th epoch.".format(epoch_i)) train_loss = 0.0 correct_cnt = 0.0 for it, train_data in enumerate(train_loader): data_dic = train_data if use_cuda: imgs, labels = Variable(data_dic['image']).cuda(), Variable(data_dic['label']).cuda() else: imgs, labels = Variable(data_dic['image']), Variable(data_dic['label']) # add channel dimension: (batch_size, D, H ,W) to (batch_size, 1, D, H ,W) # since 3D convolution requires 5D tensors img_input = imgs#.unsqueeze(1) integer_encoded = labels.data.cpu().numpy() # target should be LongTensor in loss function ground_truth = Variable(torch.from_numpy(integer_encoded)).long() if use_cuda: ground_truth = ground_truth.cuda() train_output = model(img_input) train_prob_predict = F.softmax(train_output, dim=1) _, predict = train_prob_predict.topk(1) loss = criterion(train_output, ground_truth) train_loss += loss correct_this_batch = (predict.squeeze(1) == ground_truth).sum().float() correct_cnt += correct_this_batch accuracy = float(correct_this_batch) / len(ground_truth) logging.info("batch {0} training loss is : {1:.5f}".format(it, loss.data[0])) logging.info("batch {0} training accuracy is : {1:.5f}".format(it, accuracy)) f1.write("batch {0} training loss is : {1:.5f}\n".format(it, loss.data[0])) f1.write("batch {0} training accuracy is : {1:.5f}\n".format(it, loss.data[0])) optimizer.zero_grad() loss.backward() optimizer.step() train_avg_loss = train_loss / (len(dset_train) / options.batch_size) train_avg_acu = float(correct_cnt) / len(dset_train) logging.info("Average training loss is {0:.5f} at the end of epoch {1}".format(train_avg_loss.data[0], epoch_i)) logging.info("Average training accuracy is {0:.5f} at the end of epoch {1}".format(train_avg_acu, epoch_i)) # validation -- this is a crude esitmation because there might be some paddings at the end dev_loss = 0.0 correct_cnt = 0.0 model.eval() for it, test_data in enumerate(test_loader): data_dic = test_data if use_cuda: imgs, labels = Variable(data_dic['image'], volatile=True).cuda(), Variable(data_dic['label'], volatile=True).cuda() else: imgs, labels = Variable(data_dic['image'], volatile=True), Variable(data_dic['label'], volatile=True) img_input = imgs#.unsqueeze(1) integer_encoded = labels.data.cpu().numpy() ground_truth = Variable(torch.from_numpy(integer_encoded), volatile=True).long() if use_cuda: ground_truth = ground_truth.cuda() test_output = model(img_input) test_prob_predict = F.softmax(test_output, dim=1) _, predict = test_prob_predict.topk(1) loss = criterion(test_output, ground_truth) dev_loss += loss correct_this_batch = (predict.squeeze(1) == ground_truth).sum().float() correct_cnt += (predict.squeeze(1) == ground_truth).sum() accuracy = float(correct_this_batch) / len(ground_truth) logging.info("batch {0} dev loss is : {1:.5f}".format(it, loss.data[0])) logging.info("batch {0} dev accuracy is : {1:.5f}".format(it, accuracy)) f2.write("batch {0} dev loss is : {1:.5f}\n".format(it, loss.data[0])) f2.write("batch {0} dev accuracy is : {1:.5f}\n".format(it, accuracy)) dev_avg_loss = dev_loss / (len(dset_test) / options.batch_size) dev_avg_acu = float(correct_cnt) / len(dset_test) logging.info("Average validation loss is {0:.5f} at the end of epoch {1}".format(dev_avg_loss.data[0], epoch_i)) logging.info("Average validation accuracy is {0:.5f} at the end of epoch {1}".format(dev_avg_acu, epoch_i)) if dev_avg_acu > max_acc: max_acc = dev_avg_acu max_epoch = epoch_i #if (abs(dev_avg_loss.data[0] - last_dev_loss) <= options.estop) or ((epoch_i+1)%20==0): if max_acc>=0.75: torch.save(model.state_dict(), open("3DCNN_model_" + str(epoch_i) + '_' + str(max_acc), 'wb')) last_dev_loss = dev_avg_loss.data[0] logging.info("Maximum accuracy on dev set is {0:.5f} for now".format(max_acc)) logging.info("Maximum accuracy on dev set is {0:.5f} at the end of epoch {1}".format(max_acc, max_epoch)) f1.close() f2.close() if __name__ == "__main__": ret = parser.parse_known_args() options = ret[0] if ret[1]: logging.warning("unknown arguments: {0}".format(parser.parse_known_args()[1])) main(options)